The proliferation of custom AI ASICs, or XPUs, from companies like OpenAI, Meta, and Microsoft raised questions about Nvidia's grip on the AI hardware market. However, Nvidia has invited competitors to license its intellectual property, particularly the NVLink Fusion high-speed interconnect technology.

On Monday, MediaTek became the latest to embrace NVLink Fusion for its fledgling datacenter XPU offering. In exchange, Nvidia invested $3.5 billion in convertible bonds issued by MediaTek. The GPU giant will also continue to license MediaTek's own designs for future DGX and RTX Spark systems.

NVLink Fusion, introduced early last year, started as a commercialized version of Nvidia's inter-GPU interconnect. It offers chip-to-chip connectivity for memory-coherent fabrics and switched fabrics to stitch multiple accelerators into rack-scale chips. The technology has expanded into Nvidia's blanket offering for licensing its semiconductor IP.

Amazon is already using Nvidia's MGX NVL72 reference design for its Trainium3-based racks and will adopt NVLink Fusion for Trainium4, dropping its in-house NeuronLink interconnect. By licensing NVLink Fusion, companies can focus on building competitive accelerators without worrying about scaling in production. Nvidia's MGX rack designs are part of the Open Compute Project, further reducing system design requirements.

Nvidia benefits not only from licensing fees but also from ongoing sales of its products. Initially, licensees were required to pair their XPUs with Nvidia CPUs or use Nvidia GPUs; they also needed NVSwitches for scale-up networking. While Nvidia has softened this stance—Amazon now pairs its own Graviton CPUs with Trainium—it still expects NVSwitch sales and licensing revenue.